A Diagnostic Model for Acute Lymphoblastic Leukemia Using Metaheuristics and Deep Learning Methods

Hamid Alinejad-Rokny, Mehdi Hosseinzadeh, P. Khoshaght, Samira Sadeghi, Parvaneh Asghari, Z. Arabi, Jan Lansky, P. Budinsky, Amir Masoud Rahmani, Sang-Woong LeearXiv 2024 · DOI · arXiv

Introduction

Acute Lymphoblastic Leukemia (ALL) is an aggressive hematological malignancy characterized by the rapid proliferation of immature white blood cells, or blast cells. The current diagnostic "gold standard" relies on manual microscopic examination of bone marrow and peripheral blood smears. This manual process is inherently slow, labor-intensive, and subjective, often leading to diagnostic delays or errors due to the subtle morphological similarities between healthy cells and leukemic blasts.

While deep learning has emerged as a powerful tool for medical image analysis, it faces significant hurdles in leukemia detection. Medical images are often high-dimensional and contain significant noise, leading to the "curse of dimensionality." This research addresses these challenges by proposing a hybrid model that combines deep transfer learning with advanced metaheuristic optimization to improve diagnostic accuracy and reliability.

Methodology

The researchers developed a sophisticated four-stage pipeline designed to extract high-quality features and optimize them for classification.

The proposed diagnostic pipeline, illustrating the flow from preprocessing to final classification.
  1. Preprocessing: To isolate the relevant cellular information, the model employs Otsu’s method for binary thresholding and masking. This step segments the foreground cell from the background, followed by cropping and resizing to a standard 224x224 pixel format.
  2. Feature Extraction: The study evaluated several pre-trained architectures, including ResNet, VGG19, and EfficientNetB3. DenseNet-201 was selected as the primary feature extractor due to its dense connectivity, which facilitates deep information flow and preserves fine-grained morphological features.
  3. Two-Stage Feature Selection: This "secret sauce" involves:
  4. Stage A: A Random Forest (RF) selector reduces the initial feature set to the 532 most informative attributes.
  5. Stage B: A hybrid metaheuristic algorithm, GA-BACO (Genetic Algorithm + Binary Ant Colony Optimization), further refines this subset. The GA explores the global search space, while BACO optimizes pheromone levels to identify the most discriminative features.
  6. Classification: A Multi-layer Perceptron (MLP) neural network performs the final classification into "ALL" or "HEM" (normal) categories.

Datasets

The model was trained and validated using the C-NMC 2019 dataset, which contains 15,114 images from 118 patients. This dataset is particularly challenging as it includes real-world microscopic images with varying illumination and staining noise.

Comparison of various studies, datasets, and methodologies in the field of leukemia detection.

As shown in the comparative table above, research in this field utilizes a variety of datasets and architectures. The use of the C-NMC 2019 dataset places this study alongside other high-impact works, providing a rigorous benchmark for the proposed hybrid approach.

Key Results and Insights

The DenseNet201-RF-GA-BACO-MLP model demonstrated superior performance compared to standard CNN-RNN ensembles and other state-of-the-art models:

  • Accuracy: 90.55%
  • Sensitivity (Recall): 95.94%
  • Precision: 91.07%
  • F1-Score: 93.44%

A critical finding was the dominance of DenseNet-201, which effectively mitigates vanishing gradient issues in deep networks. Furthermore, the hybrid RF-GA-BACO selection process proved essential; by curating the feature set through evolutionary and swarm intelligence, the model avoids the noise and redundancy that typically degrade performance in high-dimensional medical data. Notably, the high sensitivity (95.94%) is vital for clinical applications, as it minimizes the risk of false negatives in cancer detection.

Conclusion and Future Directions

This study presents a robust diagnostic tool that significantly aids clinicians in the automated assessment of blast cells. While the GA-BACO integration improved stability, future work will focus on enhancing the robustness of the Binary Ant Colony Optimization component. The researchers also plan to investigate hybrid deep-learning architectures and ensemble methods to further improve diagnostic metrics and ensure the model's generalizability across diverse clinical settings and staining techniques.

Summary updated

BibTeX
@misc{alinejadrokny2024diagnostic,
  author = {Hamid Alinejad-Rokny and Mehdi Hosseinzadeh and P. Khoshaght and Samira Sadeghi and Parvaneh Asghari and Z. Arabi and Jan Lansky and P. Budinsky and Amir Masoud Rahmani and Sang-Woong Lee},
  title = {{A Diagnostic Model for Acute Lymphoblastic Leukemia Using Metaheuristics and Deep Learning Methods}},
  year = {2024},
  eprint = {2406.18568},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2406.18568},
}
Figure